AGNTCY

What it is

AGNTCY is an open-source project that provides infrastructure for four things AI agents need when they cooperate with agents that were built by someone else: discovery, identity, messaging and observability.^[1] The agents it is meant to connect come from different vendors and different frameworks, so the project cannot assume a single runtime or a single model provider underneath. AGNTCY describes itself as infrastructure for an Internet of Agents, a collaboration layer that works regardless of who built an agent or where it runs.^[1]

Origin

Cisco initially open sourced AGNTCY in March 2025, with collaboration from LangChain and Galileo.^[1] On 29 July 2025 the Linux Foundation welcomed the project.^[1] Cisco, Dell Technologies, Google Cloud, Oracle and Red Hat joined as formative members, and by the time of the announcement the project had grown to more than 65 supporting companies.^[1] That list spans networking equipment, enterprise servers, cloud infrastructure and open-source software, which fits the project's premise: making agents from different stacks work together is not a problem one vendor can settle on its own.

Components

The components described when the project entered the Linux Foundation cover the basic plumbing. The Open Agent Schema Framework describes what an agent can do; identity is cryptographically verifiable, so an agent can prove which party it stands for; SLIM messaging carries messages between agents; and end-to-end observability tooling lets operators see what a multi-agent system is doing.^[1] The project's site lists six component families: an agent directory, SLIM messaging, a secure runtime, an A2A bridge for coding agents and CLIs, identity, and observability with evaluation.^[2] Put plainly, the directory is what makes an agent findable, the secure runtime gives agents a controlled place to execute, and the bridge pulls coding agents and command-line tools into the same system.

Working with A2A and MCP

AGNTCY interoperates with A2A and with Anthropic's Model Context Protocol, so that A2A agents and MCP servers can be discovered through AGNTCY directories.^[1] This is a division of labour rather than a replacement. A2A and MCP already describe how agents talk to one another and how models reach tools; AGNTCY supplies the surrounding layer — which agent this is, where to find it, how a message reaches it, and what can be observed afterwards.

Governance and reference implementation

The project is governed as a series of LF Projects, LLC.^[2] Its documentation ships a reference implementation called CoffeeAGNTCY, which demonstrates multi-agent applications built from the components.^[3] A reference application matters because it puts the directory, the identity layer and the messaging layer into one running example instead of describing each part in isolation.

Where it sits

When the agents in a system are written by different companies and built on different frameworks, connecting every pair of them separately makes integration work grow quickly with the number of agents. Discovery, identity, messaging and observability, if each team implements them again, are also hard to keep consistent. AGNTCY aims to make those four things a shared layer, and by building on A2A and MCP instead of defining rival protocols it places itself underneath them. The project went from a Cisco open-source release to a Linux Foundation project within a few months, and grew to more than 65 supporting companies.^[1]